~/australia/melbourne

I never save anything for the swim back.



← All talks

[2026]

You Can Do Everything


Transcript

Fair warning, I wasn't planning on giving a talk. The other speaker we were going to have decided to move to San Francisco and join an AI company. I found this out on Tuesday morning last week and then realized I was going to have to give a talk. But I didn't really have anything to talk about. I decided to make three different projects assuming that only maybe one of them would work, so I could have a sort of hedge.

But the spoiler is that all three projects work. I'm going to show you all three today, and I did this all within the last week. Yeah, that's kind of the general theme of this talk. But as I said, bear in mind this was very rushed, and I think I've rehearsed it one time in full; it came out to about 21 minutes. Strap in if I go a couple of minutes longer or a couple of minutes shorter; just bear with me.

Okay. My talk today is called "You Can Do Everything," and it's why I think frontier AI models are going to force us to start thinking bigger. We're not going to have an option. All of the low-hanging fruit has been picked. The golden era that we've had for the last 10 years of 0% interest rates, crazy hiring, 6-week boot camps to get, you know, $200,000 salaries for UX roles—I'm sorry to say that era is over, and we're moving into a new era.

And as I said, all the things that used to take us weeks or months, these can all be done very quickly now, and we need to adjust our ambition and our expectations for what we do in our work and also in our personal lives.

As you know, my name is Adam Brock. I'm the founder and junior vice president at Hypermatic. And I'm just going to be showing you three demos today. The first demo is going to be porting a Windows game. The second demo is going to be decompiling a Sega Dreamcast game. And the third demo is going to be building my own 3D video game. So I just want to kind of set the scene by saying this time is different.

And people often say this for various topics, but based on the curve that I've observed over the last few years—and I'm sure you have as well—if these scaling laws hold, which they seem like they will, if you look at what's kind of kicked off since September, just a month ago, where we had GPT-6 Astra and Opus 5.5 come out, it's kind of kicked off this tsunami of activity online. And you've probably seen some of these projects.

What you're looking at right now is actually a clone of Photoshop. This is not actually Photoshop; this is a clone that a couple of people pointed Claude at and basically just told it to use the real Photoshop and rebuild it using Claude. And so they've replaced every single Adobe app, every Microsoft Office app, Pro Tools, Figma, you know, all these different pieces of software. And it's just not going to kind of slow down.

And it's also kind of weirdly flooded out into the video game scene, as well, particularly with things like jailbreaking. AI basically helped to recently jailbreak the PS5 console, which means you can run any kind of software that you want on it without Sony giving you permission to do that.

And this has opened up a really weird sort of gateway into the future where you can now play PC games on your PS5 because you can run Windows games on Linux via Proton. You can basically run Windows on your PS5 and play your Steam library on your PS5, but you can also play PS5 games on a PS5 emulator on Windows running on a PS5. You've got that option as well if you prefer to do that. But it doesn't just stop at consoles; it has also made its way into video games.

Dozens of video games are being just decompiled, recompiled, you know, ported into different languages. And so this is just one quick example of this trend where I think in the future everything is just going to become everything because everything is possible.

And so this is Modern Warfare 2 rewritten in Rust and then injected with the PS3 game Skate 3's mechanics. And I'll just show you a really quick clip of this.

Anyway, you get the idea. I've been really interested in this topic. Again, I didn't really have anything to talk about, so I thought I'd just make a talk about what I've been personally interested in, hoping that you might find it interesting as well. I decided to have a go at some of these ideas. And the first idea I had was taking a Microsoft Windows game and figuring out how to get it to run in a web browser.

And so this is an old 2D multiplayer shooter game called Soldat. And it's a Windows game, so it doesn't run in the browser.

But I decided to use Codex to rewrite the source code. The cool thing about Soldat is the source code has actually been made public by the developer, so you can actually use the source code, but the source code is in Pascal, which is this other language. And so I basically got Codex to rewrite the entire source code into Rust. And that allows you to do other interesting things like port it to different platforms like the web.

And so I let Codex just go to town on it, and it just blitzed through this port. It rewrote the whole source code over to Rust. And then I basically took that Rust code and got it to compile it into WebAssembly, which allows you to run it natively in a browser. And then I also added multiplayer support as well.

I'm going to show you a quick live demo of this. But the general gist is that this is just a static HTML page with JavaScript and a WebAssembly file, and it can run in the browser. And then it's using Cloudflare Workers and Cloudflare Durable Objects to kind of track the state of the players to get the multiplayer sort of effect. So I'm just going to open this up now in my browser. And if you want to get your phone out and just open up your camera, you're welcome to join in with me if you'd like as well.

I've only tested it with like three concurrent players, so expect some lag, but you can see here I can basically play online and create a match. And then if I get my phone out, I could just scan this QR code. And you want to put your phone in landscape as well and just let that load up.

While that's loading on my phone, I'm just going to select a weapon on here. Yeah, you can see the game running in here. So I'm going to join the match and just put my microphone down for a sec. Yeah. I just killed myself in both directions. But yeah, you get the general idea. This is multiplayer Soldat. I'm just going to close this off now. That's enough. You get the idea.

We can play it again later if you'd like. Yeah, in any case, this is an example of how you can take a game from Windows or another platform, port it into Rust, and just do whatever you want with it. In this case, port it over to the web. And this took one day. This was a very tiny amount of my weekly Codex usage. And it was just using GPT-6.1 Sol on medium. So it wasn't even using the Astra frontier.

This is just the sort of second lowest model to get that done.

But I wanted to kind of take it a step further because that was kind of on easy mode because we already had the source code. It's very trivial for AI to rewrite one language to another; that's been the case for a long time. But there's this new library that's kind of gone completely insane called Reverse Engineer Anything, and I think it's up to 400,000 stars on GitHub now. And it's been out for a few weeks, and there's all these viral posts on x.com, the everything app, of course, and everyone's saying how they're decompiling Steam games in 4 seconds.

And it's amazing, and you always have to be a bit skeptical of that. I thought I should try Reverse Engineer Anything or REA for myself and figure out if it actually works. And I thought I'd kind of throw something a little bit left of field at it.

I decided to give it a Sega Dreamcast game. And so I wanted to reverse engineer Dolphin Blue, which is like this side-scrolling shoot 'em up. It's basically Metal Slug with dolphins. And I threw Claude Code at it. And let's see what it did. And so, for anyone in the room who's kind of not familiar with reverse engineering, unlike the other project we just saw where it had the source code available, there is no source code.

This is just the Dreamcast disc, essentially, with all the code sort of compiled on that disc. And so we're taking this machine code essentially and getting the AI to reverse engineer that into C code, and then understanding what that code does, and then rewriting it into readable C code so a human can look at it and actually understand what the functionality is.

This is a very difficult process to do for a human, of course, but the AI is quite good at it. And so what it does is it takes the original code and then it takes the compiled C code that it's kind of tried to guess or figure out and then it matches those up. And if it comes out the same at the other end, then it knows that's a match. It does this thousands of times.

This game has like thousands of functions. It kind of just ran through that and then it also had to code a miniature Dreamcast because the Dreamcast has its way of displaying graphics and sound. And the agent had to kind of figure out how to do all those things as well because this is not running in like an emulator. This is not like a Dreamcast emulator that we're loading a game into. This is trying to get it to run natively on other hardware.

And so I had about half a dozen AI agents all running in parallel for about 24 hours straight. And so they were doing things like handling the graphics chip, the sound chip, figuring out the code functions, naming the functions, all that sort of stuff. And yeah, long story short, it did it. It figured out how to reverse engineer that code into C. And then, because we have that source code, we can compile that down to different platforms.

Again, I wanted it to run in the browser natively. That was the first thing I tried, and I'll load that up now in the browser and just hit start on that. And because we have the source code, we can do stuff like enable widescreen, which wasn't available on the Dreamcast. We can turn on integer scaling. We can even make it run at 4K if we wanted to as well because we can increase the resolution.

This is the game running in the browser natively, and you can see it's running at 60 fps, which wasn't the case on the original hardware. And yeah, you can basically just use it as if you were playing a normal game. But I didn't want to quite stop there. So, I went back to the agent after we got this source code working and I asked it for some more stuff.

If I can get this to open up, I'll show you what we did next. Yeah, we basically got it to run in the browser, but I thought because we got the source code, you can do anything, right? I wanted to swap out the characters of the game. So I decided to make Jensen from NVIDIA the hero. And of course, the obvious enemy would be Eliezer Yudkowsky. He might be better known for writing the bestselling children's book, "If Anyone Builds It, Everyone Dies." He's kind of the poster boy for AI doomerism.

And so I decided to put them in the game. I got Codex to go through all the existing sprites in the game and regenerate these hundreds of sprites in all the different poses and stuff based on these characters. I got it to compile that down into a Mac build, so we can actually run this demo directly as a native macOS app. I'll just run that for you now. You notice it looks basically the same as our web version because it's from the same source code, with one small difference.

You can see the enemies have been swapped out, and in the top left corner, you can see we've got Jensen as the new name. So now we get Jensen, and he gets to basically ride dolphins and use his green NVIDIA machine gun and stuff like that.

Yeah. Now we can put Jensen in the game. That was pretty neat. Thank you. We should probably call my dot and congratulate that instead. But in any case, this also took one day, and I know that because it was running in parallel to the same day when I was doing the Soldat port. But this is running in Claude. This used less than 20% of my weekly Claude Ultra usage, which is kind of the highest tier.

And this was just running on Opus 5.5 medium. Again, not even using the frontier like Fable model or anything like that. 5.5 is actually a very good model, probably the best model to date. And so I was reminded of this quote from Sam Altman recently, where he was basically saying that what we expected people to do in 3 months at Y Combinator with great effort and five people is now just 17 minutes plus Codex. It's just one prompt essentially.

And his point was that we should not just do that work and stop at the 17 minutes. We should still take the entire three months and at the end of the three months ideally just have like a million times the output. That's kind of the idea. And he's trying to kind of reframe people's minds to aim a little bit bigger or dream a little bit bigger. And so I thought like a really tiny version of this would be like, okay, well let me kind of push myself a little bit and see what I could do in like a week.

And bear in mind this is just for fun in my spare time, at night and stuff. And so I thought, I wonder if I can just make a 3D game in one week and just kind of Rick Rubin my way through it. You know, I have absolutely no game dev or design experience or background or ability.

But again, I was unsure if the previous two projects would work, so I thought at least I'd have control over this and be able to show a really crappy game at the end of it if the worst thing happened. I decided to like read a couple of game design books just so I wasn't completely flying blind. And I guess the key takeaway from this book I read called "A Theory of Fun"—at the end he was really saying, as game developers, you have to be responsible with the kinds of games that you build.

And he really wanted to hammer it home that, for example, you should not build a hate crime simulator. So I decided to build a hate crime simulator. And in the sort of Zero to One Peter Thiel analogy of "What great business is nobody building?" it's like, "What fun game is nobody building?" And I think a hate crime simulator has way less competition than all of the other games.

This game is called "Office Warfare" in Korean. And it's a Korean office simulator, and you can think of it as Korean revenge cinema meets John Wick meets Office Space. And so, I got a 60-second trailer. This is real in-game footage of where I got to, just to kind of show you where it's going. And then I'll walk backwards through it to show you how I actually made each component. If you've got 60 seconds, this is going to be the trailer.

So, I'll let you enjoy. It's not safe for work, by the way.

It was just getting really good. Okay. Yeah, that was kind of where we ended up. And it's also worth noting that the in-game footage we just saw was also played, recorded, and edited by Claude as well. It played the game, recorded all the footage, and then cut it together and added the music at the end.

It's worth noting this is a full 3D environment. Of course, everything I'm going to show you is made by AI agents. I didn't design anything. I didn't code anything. I didn't use any 3D tooling myself. I just kind of acted as the Rick Rubin kind of art director in some sense. The environment was created in Blender and then imported into Godot, which is an open-source game engine. And all of the 3D models were created with the AI agents as well.

They were created in Blender through a mixture of Python scripts and also an MCP server. And then I got Codex to generate some illustrations for the characters. During the game, you've got this optional feature where all your co-workers and people are just constantly sending you messages on KakaoTalk about how it's a misdemeanor to leave your desk without asking and "come to the meeting," and all this kind of stuff. You're just kind of getting bombarded with these characters the whole time.

And then also, all the character animations were done with AI. I kind of worked with it to tell it what I wanted, and then it would iterate on it, and then I would basically say if it was good or needed improvement. It looks kind of like this. Yeah. Like, half of them are pretty retarded, but you get the idea; it's a pretty good first shot. But another thing that really helped was having an art critic sub-agent.

I realized I was just giving feedback constantly at the start, and so I'd have to kind of wait for it to finish, tell it what it needed to correct, and it just got old pretty quickly. I decided just to spin up my own sub-agent to work on my behalf, which would work with all of the other agents working. Whenever they finished a task, they would have to check with the art critic to see if it would meet my criteria or not.

And so I think it spent like half a day in one thread just working on this rain because I just would not let it get away with the rain effects that I didn't like. Yeah, I kind of iterated and finally got something that was like pretty decent, I guess. Yeah, the art critic was definitely helpful there. But I also realized I needed a background video for the main menu of the game. I decided to storyboard out a short 40-second intro video, which is just intended to be played on loop muted.

But this does have sound effects. I generated this again in GPT, storyboarded it, and then I prompted it to use Seedance, which is like the ByteDance video generation model. I created this 40-second intro video, which I'll show you in the context of the menu in a moment. But here's the kind of raw video you can have a look at. And yeah, just ignore the sound effects, basically.

Think he's looking for his hammer. Claude's a bit of a pussy, so it wouldn't let me add the blood, but you get the idea.

So yeah, also obviously you need a soundtrack. I generated 20 different industrial K-pop songs for the soundtrack and then about another two dozen event cues. Like, when you finish a level or you die, there's all these like little event cues as well. I did that with Suno AI, which is excellent for music and other sounds. And yes, this is a real game that you can actually play. I'll just quickly spin it up on the Mac because I compiled it for Mac.

I can actually just run it. If we open this up, you can see the background video playing. Oh, sorry. Sorry, one second. Oh, it's still in full screen. My bad. Okay, take two. Yeah. You can see the background video embedded now, just with some music instead. And yeah, it's all configurable, so you can turn on and off different features if you don't like the bloom effects or the depth of field stuff or like camera shaking.

You can obviously turn off the blood and gore. I don't know why you'd want to do that, but you can if you wanted to.

In any case, it starts off really easy, and then it gets gradually harder. I'll play it for like 20 seconds and then I'll just close it, and we can wrap up. Yeah. You can shoot out windows and stuff. That's pretty cool. Obviously, you need to use the shotgun for obvious reasons. What else have we got? Assault rifle. Oops. Grenades maybe. Oops. Oops. Yeah. You can do all kinds of things. You can use Uzis.

Oops. Oops. Yeah, I actually really haven't played the game that much, so I'm not very good at it. But in any case, that's kind of the idea. So, still a work in progress, obviously.

Okay. All right. Let's land this disaster, and we can wrap up and take a break. Okay. Yeah, as you saw, this is an actual game you can play. And probably the final thing I'll just mention is that obviously the AI is doing a lot of the work. And by a lot of the work, I mean all of the work. But I wanted to know just out of interest what's actually happening. I got it to create a 30-minute video tutorial for me based on all of the stuff it actually did, understanding how the code is actually structured, how the maps are procedurally generated.

Every time you play the game, the maps are totally different. And yeah, just a whole bunch of other interesting things that are in the video. That was just for my own benefit.

I'll just finish up with this final quote from Geoff Huntley. He's like a very AI-pilled person from Australia. He came up with the Ralph Loop kind of idea. He's basically living like a year in the future. Whatever he's saying, just assume that's what's going to be true in about a year. He said recently that your edge right now is understanding the possibilities and being curious. You need to be way more ambitious than you are right now, and you need to think bigger.

And I completely agree with that. In closing, I would just say I hope that you feel the AGI and make something wonderful. Thank you. And maybe don't make a hate crime simulator game.